Autor: Jakub Rusinowski · Ostatnia aktualizacja: 21 lipca 2026
Ranked for security analysis and defensive research: Log and alert triage, code review for vulnerabilities, and analysis that cannot be sent to a third party.
Best overall: AMD Ryzen AI Max+ 395
96 GB VRAM at 256 GB/s. It runs 97 of the models that qualify for this workload; the strongest is Qwen 3.7 35B-A3B at an estimated 60.3 tokens/sec.
| GPU | VRAM | MSRP | Models that fit | Best model it runs | Est. speed | |
|---|---|---|---|---|---|---|
| Best overall | AMD Ryzen AI Max+ 395 | 96 GB | $1,999 | 97 | Qwen 3.7 35B-A3B | ~60.3 tok/s |
| Best value | AMD Radeon RX 9070 XT | 16 GB | $599 | 69 | Devstral Small 2 24B | ~32.5 tok/s |
| Budget pick | Intel Arc B570 | 10 GB | $219 | 59 | Qwen 3 14B | ~27.8 tok/s |
| Most memory | NVIDIA DGX Spark | 128 GB | $4,699 | 98 | Qwen 3.7 35B-A3B | ~63.8 tok/s |
| GPU | Score | VRAM | MSRP | Models fit | Est. speed | Tok/s per watt | Cost per model |
|---|---|---|---|---|---|---|---|
| AMD Ryzen AI Max+ 395 | 81.9 | 96 GB | $1,999 | 97 | ~60.3 | 0.5 | $21 |
| NVIDIA DGX Spark | 79.2 | 128 GB | $4,699 | 98 | ~63.8 | 0.43 | $48 |
| NVIDIA RTX 6000 Ada Generation | 76.3 | 48 GB | $6,799 | 89 | ~164.8 | 0.55 | $76 |
| NVIDIA L40S | 75.8 | 48 GB | $7,499 | 89 | ~154.1 | 0.44 | $84 |
| NVIDIA GeForce RTX 5090 | 73.6 | 32 GB | $1,999 | 87 | ~59.4 | 0.1 | $23 |
| AMD Radeon RX 7900 XTX | 69.7 | 24 GB | $999 | 87 | ~33.9 | 0.1 | $11 |
| NVIDIA GeForce RTX 4090 | 69.3 | 24 GB | $1,599 | 87 | ~35.5 | 0.08 | $18 |
| NVIDIA GeForce RTX 3090 | 67 | 24 GB | $1,499 | 87 | ~33.1 | 0.09 | $17 |
| NVIDIA GeForce RTX 5070 Ti | 64.7 | 16 GB | $749 | 69 | ~39.9 | 0.13 | $11 |
| AMD Radeon RX 7900 XT | 64.2 | 20 GB | $899 | 78 | ~28.6 | 0.09 | $12 |
| NVIDIA GeForce RTX 5080 | 62.7 | 16 GB | $999 | 69 | ~42.5 | 0.12 | $14 |
| AMD Radeon RX 9070 XT | 58.2 | 16 GB | $599 | 69 | ~32.5 | 0.15 | $9 |
The AMD Ryzen AI Max+ 395 — 96 GB of VRAM runs 97 qualifying models, the strongest being Qwen 3.7.
The Intel Arc B570 at $219, which runs 59 qualifying models.
8 GB is the entry point at which a model for this workload will run at all. More memory buys a stronger model, not just a faster one.